Microsoft Turing Team
The MS Turing team develops state-of-the-art, large-scale models to solve challenging business problems across Microsoft, from Bing to Office to Xbox to Dynamics. Our mission is to expand the boundaries of natural language understanding, machine reading comprehension, question answering, transfer learning, reinforcement learning, computer vision, and even building interpretable models.
Our NLP stack includes both natural language understanding and natural language generation, in English and Universal languages. Turing NLP models are used across Microsoft to finetune for downstream tasks, such as SmartFind in Microsoft Word or Question Matching in Xbox.
For instance, the models we develop include:
- Monolingual and universal language representation
- Question answering
- Language generation
- Text prediction (e.g. Smart Compose)
- Dialog generation (e.g. bots)
Turing models power a variety of features in Bing: from search rankings to autosuggest. We deploy state of the art models to production, trained on both public and internal data.
More detailed list of tasks supported by Turing models, most of which ship to international markets:
- Ranking and retrieval
- QnA & Captions
- Answer Relevance
- Image QnA
- Direct Answers
- “People Also Ask” feature
- Ads relevance and generation
- Fast nearest neighbor search at scale
We use deep learning and machine reading comprehension (MRC) to enhance our search experience by providing direct answers to your search queries (QnA), identifying and showcasing the most relevant search results, suggesting similar content (“People Also Ask”) in lightning-fast speed!
Enterprise Semantic Search
The team develops text-based precision rankers using learned query and document encodings, allowing fast and effective search of information at the enterprise scale.
The Turing team is also developing multimodal representation learning frameworks to learn Universal language-vision representations, for tasks such as text-image QnA.
We work closely with the platform team to optimize model training for large-scale models. We also use a variety of techniques to reduce the inference latency of our models, for instance working together with the ONNX team.